Nonsense, Austen

ChatGPT and large language models are often viewed as mysterious things that run on powerful water-hungry data centres in distant deserts. But ChatGPT 2 was open source and its core functionality can be replicated in three files of Python each less than 300 lines (train, model and infer).

While most chatbots require a decent computer and graphics card to run (and training them requires renting a much more powerful computer), it's possible to run a much simpler version on your everyday computer. This game runs Andrej Karpathy's nanoGPT in-browser to generate 500 characters of Jane Austen-like prose in less than 30 seconds. It works on my MacBook Pro 2019 i5 with integrated graphics - so provided your web browser is up-to-date should work fine on your laptop too.

After play, players are invited to reflect on AI and the game. Is it neat? Educational? AI slop? Or a reflection of exactly what's wrong with AI: stealing other's work, venturing into artistic pursuits and capturing attention with a mediocre performance?

Notes

• Players are welcome to download the code and train their own models with datasets such as Shakespeare or Tolstoy. Training from scratch typically takes less than 15 minutes on a modern NVIDIA GPU.
• The game does not work on most mobiles - probably to stop people mining bitcoins or draining your battery - and so will switch to a bunch of prefetched samples in this case.
• If you have a powerful GPU, please consider giving Jan.ai a go!  Where this game is a trashy fast-food meal prepared in two minutes, Jan offers a full banquet experience with local LLMs.
• The cover image for this game is Designed by Freepik.